File size: 20,921 Bytes
5b56d44
 
 
 
 
 
 
 
 
 
 
 
83b71db
 
 
 
 
 
 
bb003e6
5b56d44
bb003e6
5b56d44
 
 
83b71db
 
 
 
 
 
 
5b56d44
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
83b71db
 
 
 
 
 
 
 
433e26f
83b71db
 
433e26f
83b71db
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5b56d44
 
 
 
 
 
 
 
 
 
 
 
 
 
bb003e6
 
83b71db
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5b56d44
 
 
 
 
 
 
 
 
 
 
 
 
83b71db
 
 
 
 
bb003e6
 
 
 
83b71db
bb003e6
5b56d44
bb003e6
 
5b56d44
 
 
 
 
 
bb003e6
83b71db
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5b56d44
 
 
 
 
 
 
 
 
 
 
 
 
 
 
83b71db
433e26f
83b71db
 
433e26f
83b71db
 
 
 
bb003e6
 
 
 
83b71db
bb003e6
83b71db
433e26f
83b71db
 
 
5b56d44
 
433e26f
 
 
 
5b56d44
 
 
 
 
 
 
 
 
 
 
 
 
83b71db
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5b56d44
 
 
 
 
 
 
 
 
 
 
 
 
83b71db
 
433e26f
 
83b71db
 
 
 
5b56d44
 
433e26f
 
 
5b56d44
 
 
 
 
 
 
 
 
 
83b71db
 
 
 
 
 
 
 
 
433e26f
83b71db
 
bb003e6
 
 
 
83b71db
 
 
 
 
 
 
 
 
 
 
bb003e6
83b71db
5b56d44
 
 
 
bb003e6
5b56d44
 
 
 
 
bb003e6
5b56d44
 
 
 
83b71db
 
 
 
 
 
5b56d44
83b71db
 
 
5b56d44
 
 
 
 
 
 
 
83b71db
 
 
 
 
 
 
 
5b56d44
83b71db
 
 
 
 
433e26f
 
 
 
83b71db
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
433e26f
83b71db
 
 
 
 
 
 
 
 
 
5b56d44
 
 
 
 
 
 
 
 
 
 
 
 
bb003e6
 
 
 
 
83b71db
 
 
 
bb003e6
 
 
 
 
 
83b71db
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
bb003e6
83b71db
5b56d44
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
83b71db
 
 
 
 
 
 
 
 
 
 
5b56d44
83b71db
5b56d44
83b71db
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
433e26f
83b71db
433e26f
83b71db
 
 
 
 
 
433e26f
 
 
83b71db
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
"""Post-processing pipeline for photorealistic face output.

Neural net components:
- CodeFormer (primary): face restoration with controllable fidelity-quality tradeoff
- GFPGAN (fallback): face restoration for diffusion artifact repair
- Real-ESRGAN: neural super-resolution for background regions
- ArcFace: identity verification to flag drift between input/output

Classical components:
- Multi-band Laplacian pyramid blending (replaces simple alpha blend)
- Frequency-aware sharpening (recovers fine skin texture)
- Color histogram matching (ensures skin tone consistency)
"""

from __future__ import annotations

import cv2
import numpy as np

# Singleton model caches -- load once, reuse across calls
_CODEFORMER_MODEL = None
_GFPGAN_HELPER = None
_REALESRGAN_UPSAMPLER = None
_ARCFACE_APP = None


def laplacian_pyramid_blend(
    source: np.ndarray,
    target: np.ndarray,
    mask: np.ndarray,
    levels: int = 6,
) -> np.ndarray:
    """Multi-band Laplacian pyramid blending for seamless compositing.

    Unlike simple alpha blending which creates visible halos at mask edges,
    Laplacian blending operates at multiple frequency bands. Low frequencies
    (overall color/lighting) blend smoothly, high frequencies (skin texture,
    pores, hair) transition sharply. This eliminates the "pasted on" look.

    Args:
        source: BGR image to blend IN (the surgical result).
        target: BGR image to blend INTO (the original photo).
        mask: Float32 mask [0-1] (1 = source region).
        levels: Number of pyramid levels (6 works well for 512x512).

    Returns:
        Seamlessly composited BGR image.
    """
    # Ensure same size
    h, w = target.shape[:2]
    source = cv2.resize(source, (w, h)) if source.shape[:2] != (h, w) else source

    # Normalize mask
    mask_f = mask.astype(np.float32)
    if mask_f.max() > 1.0:
        mask_f = mask_f / 255.0
    mask_3ch = np.stack([mask_f] * 3, axis=-1) if mask_f.ndim == 2 else mask_f

    # Make dimensions divisible by 2^levels
    factor = 2**levels
    new_h = (h + factor - 1) // factor * factor
    new_w = (w + factor - 1) // factor * factor

    if new_h != h or new_w != w:
        source = cv2.resize(source, (new_w, new_h))
        target = cv2.resize(target, (new_w, new_h))
        mask_3ch = cv2.resize(mask_3ch, (new_w, new_h))

    src_f = source.astype(np.float32)
    tgt_f = target.astype(np.float32)

    # Build Gaussian pyramids for the mask
    mask_pyr = [mask_3ch]
    for _ in range(levels):
        mask_pyr.append(cv2.pyrDown(mask_pyr[-1]))

    # Build Laplacian pyramids for source and target
    src_lap = _build_laplacian_pyramid(src_f, levels)
    tgt_lap = _build_laplacian_pyramid(tgt_f, levels)

    # Blend each level using the mask at that resolution
    blended_lap = []
    for i in range(levels + 1):
        sl = src_lap[i]
        tl = tgt_lap[i]
        ml = mask_pyr[i]
        # Resize mask to match level shape if needed
        if ml.shape[:2] != sl.shape[:2]:
            ml = cv2.resize(ml, (sl.shape[1], sl.shape[0]))
        blended = sl * ml + tl * (1.0 - ml)
        blended_lap.append(blended)

    # Reconstruct from blended Laplacian
    result = _reconstruct_from_laplacian(blended_lap)

    # Crop back to original size
    result = result[:h, :w]
    return np.clip(result, 0, 255).astype(np.uint8)


def _build_laplacian_pyramid(
    image: np.ndarray,
    levels: int,
) -> list[np.ndarray]:
    """Build Laplacian pyramid from an image."""
    gaussian = [image.copy()]
    for _ in range(levels):
        gaussian.append(cv2.pyrDown(gaussian[-1]))

    laplacian = []
    for i in range(levels):
        upsampled = cv2.pyrUp(gaussian[i + 1])
        # Match sizes (pyrUp can add a pixel)
        gh, gw = gaussian[i].shape[:2]
        upsampled = upsampled[:gh, :gw]
        laplacian.append(gaussian[i] - upsampled)

    laplacian.append(gaussian[-1])  # coarsest level
    return laplacian


def _reconstruct_from_laplacian(pyramid: list[np.ndarray]) -> np.ndarray:
    """Reconstruct image from Laplacian pyramid."""
    image = pyramid[-1].copy()
    for i in range(len(pyramid) - 2, -1, -1):
        image = cv2.pyrUp(image)
        lh, lw = pyramid[i].shape[:2]
        image = image[:lh, :lw]
        image = image + pyramid[i]
    return image


def frequency_aware_sharpen(
    image: np.ndarray,
    strength: float = 0.3,
    radius: int = 3,
) -> np.ndarray:
    """Sharpen high-frequency detail (skin texture, pores) without amplifying noise.

    Uses unsharp masking in LAB space (luminance only) to avoid
    color fringing. Preserves the smooth look of diffusion output
    while recovering fine texture detail.

    Args:
        image: BGR image.
        strength: Sharpening strength (0.2-0.5 typical for faces).
        radius: Gaussian blur radius for unsharp mask.

    Returns:
        Sharpened BGR image.
    """
    image_u8 = np.clip(image, 0, 255).astype(np.uint8)
    lab = cv2.cvtColor(image_u8, cv2.COLOR_BGR2LAB).astype(np.float32)
    l_channel = lab[:, :, 0]

    # Unsharp mask on luminance only
    ksize = radius * 2 + 1
    blurred = cv2.GaussianBlur(l_channel, (ksize, ksize), 0)
    sharpened = l_channel + strength * (l_channel - blurred)

    lab[:, :, 0] = np.clip(sharpened, 0, 255)
    return cv2.cvtColor(lab.astype(np.uint8), cv2.COLOR_LAB2BGR)


def restore_face_gfpgan(
    image: np.ndarray,
    upscale: int = 1,
) -> np.ndarray:
    """Restore face quality using GFPGAN.

    Fixes common diffusion artifacts: blurry eyes, distorted features,
    inconsistent skin texture. The restored face is then blended back
    into the original for a natural look.

    Args:
        image: BGR face image (any size).
        upscale: Upscale factor (1 = same size, 2 = 2x).

    Returns:
        Restored BGR image, or original if GFPGAN unavailable.
    """
    try:
        from gfpgan import GFPGANer
    except ImportError:
        return image

    # GFPGAN requires 3-channel BGR input
    if image.ndim == 2 or (image.ndim == 3 and image.shape[2] == 1):
        image = cv2.cvtColor(image, cv2.COLOR_GRAY2BGR)

    try:
        global _GFPGAN_HELPER
        # Singleton: avoid reloading ~300MB GFPGAN model on every call
        if _GFPGAN_HELPER is None:
            _GFPGAN_HELPER = GFPGANer(
                model_path="https://github.com/TencentARC/GFPGAN/releases/download/v1.3.0/GFPGANv1.3.pth",
                upscale=upscale,
                arch="clean",
                channel_multiplier=2,
                bg_upsampler=None,
            )
        _, _, restored = _GFPGAN_HELPER.enhance(
            image,
            has_aligned=False,
            only_center_face=True,
            paste_back=True,
        )
        if restored is not None:
            return restored
    except Exception:
        pass

    return image


def restore_face_codeformer(
    image: np.ndarray,
    fidelity: float = 0.7,
    upscale: int = 1,
) -> np.ndarray:
    """Restore face quality using CodeFormer (neural net).

    CodeFormer uses a Transformer-based codebook lookup to restore degraded
    faces. The fidelity parameter controls the quality-fidelity tradeoff:
    lower values produce higher quality but may alter identity slightly,
    higher values preserve identity but fix fewer artifacts.

    Args:
        image: BGR face image.
        fidelity: Quality-fidelity balance (0.0=quality, 1.0=fidelity). 0.7 default.
        upscale: Upscale factor (1 = same size).

    Returns:
        Restored BGR image, or original if CodeFormer unavailable.
    """
    try:
        import torch
        from codeformer.basicsr.utils import img2tensor, tensor2img
        from codeformer.basicsr.utils.download_util import load_file_from_url
        from codeformer.facelib.utils.face_restoration_helper import FaceRestoreHelper
        from torchvision.transforms.functional import normalize as tv_normalize
    except ImportError:
        return image

    # CodeFormer requires 3-channel BGR input
    if image.ndim == 2 or (image.ndim == 3 and image.shape[2] == 1):
        image = cv2.cvtColor(image, cv2.COLOR_GRAY2BGR)

    try:
        global _CODEFORMER_MODEL
        from codeformer.basicsr.archs.codeformer_arch import CodeFormer as CodeFormerArch
        from codeformer.inference_codeformer import set_realesrgan as _unused  # noqa: F401

        device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

        if _CODEFORMER_MODEL is None:
            model = CodeFormerArch(
                dim_embd=512,
                codebook_size=1024,
                n_head=8,
                n_layers=9,
                connect_list=["32", "64", "128", "256"],
            ).to(device)

            ckpt_path = load_file_from_url(
                url="https://github.com/sczhou/CodeFormer/releases/download/v0.1.0/codeformer.pth",
                model_dir="weights/CodeFormer",
                progress=True,
            )
            checkpoint = torch.load(ckpt_path, map_location=device, weights_only=True)
            model.load_state_dict(checkpoint["params_ema"])
            model.eval()
            _CODEFORMER_MODEL = model
        model = _CODEFORMER_MODEL

        face_helper = FaceRestoreHelper(
            upscale,
            face_size=512,
            crop_ratio=(1, 1),
            det_model="retinaface_resnet50",
            save_ext="png",
            device=device,
        )
        face_helper.read_image(image)
        face_helper.get_face_landmarks_5(only_center_face=True)
        face_helper.align_warp_face()

        for cropped_face in face_helper.cropped_faces:
            face_t = img2tensor(cropped_face / 255.0, bgr2rgb=True, float32=True)
            tv_normalize(face_t, (0.5, 0.5, 0.5), (0.5, 0.5, 0.5), inplace=True)
            face_t = face_t.unsqueeze(0).to(device)

            with torch.no_grad():
                output = model(face_t, w=fidelity, adain=True)[0]
                restored = tensor2img(output, rgb2bgr=True, min_max=(-1, 1))
            restored = restored.astype(np.uint8)
            face_helper.add_restored_face(restored)

        face_helper.get_inverse_affine(None)
        restored_img = face_helper.paste_faces_to_image()
        if restored_img is not None:
            return restored_img
    except Exception:
        pass

    return image


def enhance_background_realesrgan(
    image: np.ndarray,
    mask: np.ndarray,
    outscale: int = 2,
) -> np.ndarray:
    """Enhance non-face background regions using Real-ESRGAN neural upscaler.

    Only applies to regions outside the surgical mask to improve overall
    image quality without interfering with the face restoration pipeline.

    Args:
        image: BGR image.
        mask: Float32 mask [0-1] where 1 = face region (skip these pixels).
        outscale: Upscale factor (2 = 2x resolution, then downsample back).

    Returns:
        Enhanced BGR image at original resolution.
    """
    try:
        import torch
        from basicsr.archs.rrdbnet_arch import RRDBNet
        from realesrgan import RealESRGANer
    except ImportError:
        return image

    try:
        global _REALESRGAN_UPSAMPLER
        if _REALESRGAN_UPSAMPLER is None:
            model = RRDBNet(
                num_in_ch=3, num_out_ch=3, num_feat=64, num_block=23, num_grow_ch=32, scale=4
            )
            _REALESRGAN_UPSAMPLER = RealESRGANer(
                scale=4,
                model_path="https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.0/RealESRGAN_x4plus.pth",
                model=model,
                tile=400,
                tile_pad=10,
                pre_pad=0,
                half=torch.cuda.is_available(),
            )
        enhanced, _ = _REALESRGAN_UPSAMPLER.enhance(image, outscale=outscale)

        # Downscale back to original size
        h, w = image.shape[:2]
        enhanced = cv2.resize(enhanced, (w, h), interpolation=cv2.INTER_LANCZOS4)

        # Only apply enhancement to background (outside mask)
        mask_f = mask.astype(np.float32)
        if mask_f.max() > 1.0:
            mask_f /= 255.0
        mask_3ch = np.stack([mask_f] * 3, axis=-1) if mask_f.ndim == 2 else mask_f

        # Keep face region from original, use enhanced for background
        result = np.clip(
            image.astype(np.float32) * mask_3ch + enhanced.astype(np.float32) * (1.0 - mask_3ch),
            0,
            255,
        ).astype(np.uint8)
        return result
    except Exception:
        pass

    return image


def verify_identity_arcface(
    original: np.ndarray,
    result: np.ndarray,
    threshold: float = 0.5,
) -> dict:
    """Verify output preserves input identity using ArcFace neural net.

    Computes cosine similarity between ArcFace embeddings of the original
    and result images. If similarity drops below threshold, flags identity
    drift -- meaning the postprocessing or diffusion altered the person's
    appearance too much.

    Args:
        original: BGR original face image.
        result: BGR post-processed output image.
        threshold: Minimum cosine similarity to pass (0.5 = same person).

    Returns:
        Dict with 'similarity' (float), 'passed' (bool), 'message' (str).
    """
    try:
        from insightface.app import FaceAnalysis
    except ImportError:
        return {
            "similarity": -1.0,
            "passed": True,
            "message": "InsightFace not installed — identity check skipped",
        }

    try:
        global _ARCFACE_APP
        if _ARCFACE_APP is None:
            _ARCFACE_APP = FaceAnalysis(
                name="buffalo_l",
                providers=["CUDAExecutionProvider", "CPUExecutionProvider"],
            )
            _ARCFACE_APP.prepare(ctx_id=0 if _has_cuda() else -1, det_size=(320, 320))
        app = _ARCFACE_APP

        orig_faces = app.get(original)
        result_faces = app.get(result)

        if not orig_faces or not result_faces:
            return {
                "similarity": -1.0,
                "passed": True,
                "message": "Could not detect face in one/both images — check skipped",
            }

        orig_emb = orig_faces[0].embedding
        result_emb = result_faces[0].embedding

        sim = float(
            np.dot(orig_emb, result_emb)
            / (np.linalg.norm(orig_emb) * np.linalg.norm(result_emb) + 1e-8)
        )
        sim = float(np.clip(sim, 0, 1))

        passed = sim >= threshold
        if passed:
            msg = f"Identity preserved (similarity={sim:.3f})"
        else:
            msg = f"WARNING: Identity drift detected (similarity={sim:.3f} < {threshold})"

        return {"similarity": sim, "passed": passed, "message": msg}
    except Exception as e:
        return {
            "similarity": -1.0,
            "passed": True,
            "message": f"Identity check failed: {e}",
        }


def _has_cuda() -> bool:
    try:
        import torch

        return torch.cuda.is_available()
    except ImportError:
        return False


def histogram_match_skin(
    source: np.ndarray,
    reference: np.ndarray,
    mask: np.ndarray,
) -> np.ndarray:
    """Match skin color histogram of source to reference within masked region.

    More robust than simple mean/std matching — preserves the full
    distribution of skin tones including highlights and shadows.

    Args:
        source: BGR image whose skin tone to adjust.
        reference: BGR image with target skin tone.
        mask: Float32 mask [0-1] of skin region.

    Returns:
        Color-matched BGR image.
    """
    # Ensure 2D mask for per-channel indexing
    m = mask
    if m.ndim == 3:
        m = m[:, :, 0]
    mask_bool = m > 0.3 if m.dtype == np.float32 else m > 76

    if not np.any(mask_bool):
        return source

    # Clip to valid uint8 range before LAB conversion to prevent overflow
    # on images with saturated or out-of-range pixel values
    src_u8 = np.clip(source, 0, 255).astype(np.uint8)
    ref_u8 = np.clip(reference, 0, 255).astype(np.uint8)
    src_lab = cv2.cvtColor(src_u8, cv2.COLOR_BGR2LAB).astype(np.float32)
    ref_lab = cv2.cvtColor(ref_u8, cv2.COLOR_BGR2LAB).astype(np.float32)

    for ch in range(3):
        src_vals = src_lab[:, :, ch][mask_bool]
        ref_vals = ref_lab[:, :, ch][mask_bool]

        if len(src_vals) == 0 or len(ref_vals) == 0:
            continue

        # CDF matching
        src_sorted = np.sort(src_vals)
        ref_sorted = np.sort(ref_vals)

        # Interpolate reference CDF to match source length
        src_cdf = np.linspace(0, 1, len(src_sorted))
        ref_cdf = np.linspace(0, 1, len(ref_sorted))

        # Map source values through reference distribution
        mapping = np.interp(src_cdf, ref_cdf, ref_sorted)

        # Create lookup from source intensity to matched intensity
        src_flat = src_lab[:, :, ch].ravel()
        matched = np.interp(src_flat, src_sorted, mapping)
        matched_2d = matched.reshape(src_lab.shape[:2])

        # Apply only in mask region
        src_lab[:, :, ch] = np.where(mask_bool, matched_2d, src_lab[:, :, ch])

    result_lab = np.clip(src_lab, 0, 255).astype(np.uint8)
    return cv2.cvtColor(result_lab, cv2.COLOR_LAB2BGR)


def full_postprocess(
    generated: np.ndarray,
    original: np.ndarray,
    mask: np.ndarray,
    restore_mode: str = "codeformer",
    codeformer_fidelity: float = 0.7,
    use_realesrgan: bool = True,
    use_laplacian_blend: bool = True,
    sharpen_strength: float = 0.25,
    verify_identity: bool = True,
    identity_threshold: float = 0.5,
) -> dict:
    """Full neural net + classical post-processing pipeline for maximum photorealism.

    Pipeline:
    1. Face restoration: CodeFormer (primary) or GFPGAN (fallback) neural nets
    2. Background enhancement: Real-ESRGAN neural upscaler (non-face regions)
    3. Skin tone histogram matching to original (classical)
    4. Frequency-aware sharpening for texture recovery (classical)
    5. Laplacian pyramid blending for seamless compositing (classical)
    6. ArcFace identity verification (neural net quality gate)

    Args:
        generated: BGR generated/warped face image.
        original: BGR original face image.
        mask: Float32 surgical mask [0-1].
        restore_mode: 'codeformer', 'gfpgan', or 'none'.
        codeformer_fidelity: CodeFormer fidelity weight (0=quality, 1=fidelity).
        use_realesrgan: Apply Real-ESRGAN to background regions.
        use_laplacian_blend: Use Laplacian blend vs simple alpha blend.
        sharpen_strength: Texture sharpening amount (0 = none).
        verify_identity: Run ArcFace identity check at the end.
        identity_threshold: Min cosine similarity to pass identity check.

    Returns:
        Dict with 'image' (composited BGR), 'identity_check' (dict), 'restore_used' (str).
    """
    result = generated.copy()
    restore_used = "none"

    # Step 1: Neural face restoration (CodeFormer > GFPGAN > skip)
    if restore_mode == "codeformer":
        restored = restore_face_codeformer(result, fidelity=codeformer_fidelity)
        if restored is not result:
            result = restored
            restore_used = "codeformer"
        else:
            # CodeFormer unavailable, fall back to GFPGAN
            pre_gfpgan = result
            result = restore_face_gfpgan(result)
            restore_used = "gfpgan" if result is not pre_gfpgan else "none"
    elif restore_mode == "gfpgan":
        restored = restore_face_gfpgan(result)
        if restored is not result:
            result = restored
            restore_used = "gfpgan"

    # Step 2: Neural background enhancement
    if use_realesrgan:
        result = enhance_background_realesrgan(result, mask)

    # Step 3: Skin tone histogram matching (classical)
    result = histogram_match_skin(result, original, mask)

    # Step 4: Sharpen texture (classical)
    if sharpen_strength > 0:
        result = frequency_aware_sharpen(result, strength=sharpen_strength)

    # Step 5: Blend into original (classical)
    if use_laplacian_blend:
        composited = laplacian_pyramid_blend(result, original, mask)
    else:
        mask_f = mask.astype(np.float32)
        if mask_f.max() > 1.0:
            mask_f /= 255.0
        mask_3ch = np.stack([mask_f] * 3, axis=-1) if mask_f.ndim == 2 else mask_f
        composited = (
            result.astype(np.float32) * mask_3ch + original.astype(np.float32) * (1.0 - mask_3ch)
        ).astype(np.uint8)

    # Step 6: Neural identity verification
    identity_check = {"similarity": -1.0, "passed": True, "message": "skipped"}
    if verify_identity:
        identity_check = verify_identity_arcface(
            original,
            composited,
            threshold=identity_threshold,
        )

    return {
        "image": composited,
        "identity_check": identity_check,
        "restore_used": restore_used,
    }